From mboxrd@z Thu Jan 1 00:00:00 1970 From: Eric Dumazet Subject: Re: [PATCH net-next] tcp: increase throughput when reordering is high Date: Wed, 21 Aug 2013 20:52:28 -0700 Message-ID: <1377143548.4226.134.camel@edumazet-glaptop> References: <1377131363-31341-1-git-send-email-ycheng@google.com> Mime-Version: 1.0 Content-Type: text/plain; charset="UTF-8" Content-Transfer-Encoding: 7bit Cc: davem@davemloft.net, ncardwell@google.com, edumazet@google.com, netdev@vger.kernel.org To: Yuchung Cheng Return-path: Received: from mail-pa0-f49.google.com ([209.85.220.49]:63590 "EHLO mail-pa0-f49.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752506Ab3HVDwa (ORCPT ); Wed, 21 Aug 2013 23:52:30 -0400 Received: by mail-pa0-f49.google.com with SMTP id ld10so1582396pab.8 for ; Wed, 21 Aug 2013 20:52:29 -0700 (PDT) In-Reply-To: <1377131363-31341-1-git-send-email-ycheng@google.com> Sender: netdev-owner@vger.kernel.org List-ID: On Wed, 2013-08-21 at 17:29 -0700, Yuchung Cheng wrote: > The stack currently detects reordering and avoid spurious > retransmission very well. However the throughput is sub-optimal under > high reordering because cwnd is increased only if the data is deliverd > in order. I.e., FLAG_DATA_ACKED check in tcp_ack(). The more packet > are reordered the worse the throughput is. > > Therefore when reordering is proven high, cwnd should advance whenever > the data is delivered regardless of its ordering. If reordering is low, > conservatively advance cwnd only on ordered deliveries in Open state, > and retain cwnd in Disordered state (RFC5681). > > Using netperf on a qdisc setup of 20Mbps BW and random RTT from 45ms > to 55ms (for reordering effect). This change increases TCP throughput > by 20 - 25% to near bottleneck BW. > > A special case is the stretched ACK with new SACK and/or ECE mark. > For example, a receiver may receive an out of order or ECN packet with > unacked data buffered because of LRO or delayed ACK. The principle on > such an ACK is to advance cwnd on the cummulative acked part first, > then reduce cwnd in tcp_fastretrans_alert(). > > Signed-off-by: Yuchung Cheng > --- > net/ipv4/tcp_input.c | 32 ++++++++++++++++++++------------ > 1 file changed, 20 insertions(+), 12 deletions(-) Acked-by: Eric Dumazet